go/mysql: Fix MariadbGTIDSet multi-domain support.

[enisoc]

This fixes two bugs:

1. AddGTID was mutating the original set. GTIDSet implementations are
   supposed to be immutable.

2. AddGTID and parseMariadbGTIDSet did not enforce any order among
   domains, yet Equal and other functions expected order to be enforced.

This switches from a list to a map to better represent the fact that the order of domains doesn't matter, which also simplifies some code. We always enforce order when serializing back to a string.

[enisoc]

@deepthi I got the tests fixed so this is ready for review now.

@PrismaPhonic I can't add you as a reviewer, but I'd appreciate if you can take a look.

[deepthi]

Is the Domain monotonically increasing in the Position string we get from MariaDB? If we parse that string into GTID and convert back to string, do we get back the same value?

[enisoc]

Is the Domain monotonically increasing in the Position string we get from MariaDB?

It looks that way in the examples I've seen in the docs, but I haven't found any documented guarantee that it's always the case.

If we parse that string into GTID and convert back to string, do we get back the same value?

Given the above, we can't be sure, short of checking the MariaDB code. Are you thinking of a case in which it might be a problem to send back a sorted value when given an out-of-order value?

[deepthi]

Given the above, we can't be sure, short of checking the MariaDB code. Are you thinking of a case in which it might be a problem to send back a sorted value when given an out-of-order value?

Right. Anywhere where we do replicate until semantics and pass in a stop position.
Let us assume MariaDB hands them out sorted and we preserve the sorting. We are still better off than the previous implemetation.

[deepthi]

LGTM

[enisoc]

Anywhere where we do replicate until semantics and pass in a stop position.

Do you think it would be a problem to hand back the domains in a different order? I would assume MariaDB internally parses and/or compares them in such a way that order doesn't matter, though I haven't found documented proof of this.

[deepthi]

Do you think it would be a problem to hand back the domains in a different order? I would assume MariaDB internally parses and/or compares them in such a way that order doesn't matter, though I haven't found documented proof of this.

I don't know. We could test it by running a vitess cluster on mariadb.